Physics-informed structural health monitoring with reinforcement learning and autonomous inspection tasking

A method for real-time structural monitoring of a physical structure, the method comprising:

• receiving time-stamped signals from a strain gauge, a tri-axial accelerometer, and a temperature sensor mounted on the structure;
• Gaussian-filtering and z-score normalizing the signals and extracting a frequency-domain feature vector by fast Fourier transform (FFT);
• inputting the feature vector to a physics-informed neural-network surrogate structural model and retraining the surrogate in real time using stochastic-gradient-descent and/or back-propagation with newly received feature vectors;
• generating an anomaly score by computing a residual between an output of the surrogate structural model and the frequency-domain feature vector;
• adapting, based on monitoring and inspection feedback and according to a reinforcement-learning policy, both (i) a sampling interval of the sensor network and (ii) a threshold applied to the anomaly score;
• and
• generating a monitoring output identifying regions whose anomaly scores exceed the threshold.

A system for real-time structural monitoring of a physical structure, the system comprising:

• a sensor suite mounted on the structure comprising a strain gauge, a tri-axial accelerometer, and a temperature sensor;
• an edge processor disposed on the structure and a cloud analytics server communicatively coupled to the edge processor, the edge processor and the cloud analytics server collectively configured to:
• (i) receive time-stamped signals from the sensor suite, Gaussian-filter the signals, z-score normalize the filtered signals, and compute a fast Fourier transform (FFT) to obtain a frequency-domain feature vector;
• (ii) input the frequency-domain feature vector and a temperature as boundary-condition input to a physics-informed neural-network surrogate structural model, and retrain the surrogate in real time using stochastic-gradient-descent and/or back-propagation with newly received feature vectors;
• (iii) compute, in real time, an anomaly score by computing a residual between an output of the surrogate structural model and the frequency-domain feature vector, and adapt, according to a reinforcement-learning policy trained on monitoring and inspection feedback, both (A) a sampling interval for the sensor suite and (B) a threshold applied to the anomaly score;
• and
• (iv) generate a monitoring output identifying regions whose anomaly scores exceed the threshold.